Two-level HLM

Hi, I have a question regarding a two-level HLM model, and I am not sure if my designation of the level 1 and level 2 predictors is methodologically sound.

My data consisted of (1) teachers’ perceptions of different components of school leadership (i.e., the leadership practices of their principals), (2) teachers’ perceptions of different components of their own practices and (3) teachers’ perceptions of their own competencies. The data was collected from the same sample of teachers. Therefore, demographic-wise, only teachers are sampled.

My intention is to find out which of the school leadership components and teacher practice components predict teachers’ competencies.

Is it okay if my HLM model is as follows:

  • Outcome variable: Teachers’ perceptions of their own competencies
  • Level 1 predictors: Teachers’ perceptions of different components of their own practices
  • Level 2 predictors: Teachers’ perceptions of different components of school leadership (aggregated by school; across schools)

Basically, what I am doing is using only teachers’ perceptions but on three different areas. Is it okay?

Thank you.

if I understand your intentions correctly, I would say there is no modelling reason to not analyze the data the way you propose.

However, interpreting the results will be very tricky. One problem I see is that I would expect both self-reported competencies and perceptions of practices and leadership to be largely influenced by personality traits (e.g. the tendency to be humble, critical, …). This means that

  1. your primary outcome might have limited accuracy: for example, the least competent may tend to overestimate their competency (Dunning-Kruger effect)
  2. personality may add spurious correlations: for example a person that is negative/critical will put a low ranking for both their competence and school leadership.

One thing you can do in this situation is treat the data as having measurement error. But unless you can correct for biases with enough data points (multiple teachers coding each administrator and many administrators coded per teacher), this may not help much.

I’m not sure what you mean by the levels. In an education setting, a hierarchical model might have students nested in classrooms, nested in schools, nested in school districts, nested in states. Each of those could be considered a “level”, but “level” is overloaded in statistics to also refer to a vaue in a group. I don’t see how the level 1 and level 2 predictors relate to each other, but it’s not in the nesting sense, is it?

Thank you both for the comments.

Regarding the point on self-perceptions “largely influenced by personality traits”, there are sufficient data points (as Bob rightly pointed out) to correct for it.

As for the point about nesting, this is what I am unsure of. Typically, a hierarchical model in this scenario would consist of teachers nested in schools (i.e., principals) where teachers are considered “level 1” and schools are considered “level 2”. The assumption, regarding leadership practices, I am making is that teachers’ perceptions of principals’ leadership practices is more valid than principals’ self-reported perceptions of their own leadership practices. Hence, my intention to use teachers’ perceptions of principals’ practices.

Therefore my concern is: Could teachers’ perceptions of principals’ practices be considered as information about the context like how principals’ self-perceptions can be treated as information about the context (of their respective schools, that is) within which teachers are nested in?

I’m not sure what you mean by “context” in “information about the context”, but they should certainly provide some information. My concern with any self reporting like this would be the bias. I’d be more comfortable if multiple teachers report on principals, for example.

But this wouldn’t be a hierarchical model so much as another predictor.